This document discusses query modeling using non-relevance information for relevance feedback. It presents a model that estimates a relevant model from relevant documents and a non-relevant model from non-relevant documents. When expanding the initial query, terms are weighted based on their probability in the relevant model normalized by their probability in the non-relevant model. Experiments on the TREC 2008 Relevance Feedback Track optimize parameters and submit runs using the top 10 terms from the expanded query model.

1. The University of Amsterdam at the
TREC 2008 Relevance Feedback Track
Query Modeling Using Non-relevance Information
Edgar Meij, W. Weerkamp, J. He, and M. de Rijke
ISLA
University of Amsterdam
http://ilps.science.uva.nl
TREC 2008

2. Introduction Model Experiments Conclusion
Outline
Introduction
Model
Experiments
Conclusion

3. Introduction Model Experiments Conclusion
Motivation
• Pseudo-relevance feedback approaches generally assume
a term’s non-relevance status is implicitly indicated by its
absence
• How should we interpret explicit non-relevance information
in a generative language modeling setting?

4. Introduction Model Experiments Conclusion
Retrieval Model
• Documents are ranked according to the KL-divergence
between a query model and each document model
P(t|θQ )
Score(D,Q) = − P(t|θQ ) log
P(t|θD )
t∈V
rank
= − P(t|θQ ) log P(t|θD )
t∈V
• Document models are smoothed using a reference corpus
• We use Jelinek-Mercer smoothing
P(t|θD ) = (1 − λD )P(t|D) + λD P(t)

5. Introduction Model Experiments Conclusion
Retrieval Model
• Documents are ranked according to the KL-divergence
between a query model and each document model
P(t|θQ )
Score(D,Q) = − P(t|θQ ) log
P(t|θD )
t∈V
rank
= − P(t|θQ ) log P(t|θD )
t∈V
• Document models are smoothed using a reference corpus
• We use Jelinek-Mercer smoothing
P(t|θD ) = (1 − λD )P(t|D) + λD P(t)

6. Introduction Model Experiments Conclusion
Query Modeling
• Assumption: the better the query model reflects the
information need, the better the results
• Baseline: Each query term is equally important and
receives an equal probability mass (set A)
c(t, Q)
P(t|θQ ) = P(t|Q) =
|Q|
• Cast pseudo-relevance feedback as query model updating
ˆ
P(t|θQ ) = (1 − λQ )P(t|Q) + λQ P(t|θQ )
• Smooth the initial query by adding and (re)weighing terms

7. Introduction Model Experiments Conclusion
Query Modeling
• Assumption: the better the query model reflects the
information need, the better the results
• Baseline: Each query term is equally important and
receives an equal probability mass (set A)
c(t, Q)
P(t|θQ ) = P(t|Q) =
|Q|
• Cast pseudo-relevance feedback as query model updating
ˆ
P(t|θQ ) = (1 − λQ )P(t|Q) + λQ P(t|θQ )
• Smooth the initial query by adding and (re)weighing terms

8. Introduction Model Experiments Conclusion
Outline
Introduction
Model
Experiments
Conclusion

9. Introduction Model Experiments Conclusion
(Non) Relevant Models
• Relevant model estimated using interpolated MLE on the
set of relevant documents:
P(t|θR ) = δ1 P(t) + (1 − δ1 )P(t|R)
D∈R P(t|D)
= δ1 P(t) + (1 − δ1 )
|R|
• Non-relevant model likewise:
P(t|θ¬R ) = δ2 P(t) + (1 − δ2 )P(t|¬R)
P(t|D)
= δ2 P(t) + (1 − δ2 ) D∈¬R
|¬R|

10. Introduction Model Experiments Conclusion
Our Model
ˆ
In order to arrive at an expanded query model θQ , we sample
terms proportional to the following:
• Each term is sampled according to the probability of
observing that term in each relevant document
• For each relevant document, adjust the probability mass of
each term by
• the probability of occurring given the relevant model
• normalized by its probability given the non-relevant model

11. Introduction Model Experiments Conclusion
Our Model
ˆ
In order to arrive at an expanded query model θQ , we sample
terms proportional to the following:
• Each term is sampled according to the probability of
observing that term in each relevant document
• For each relevant document, adjust the probability mass of
each term by
• the probability of occurring given the relevant model
• normalized by its probability given the non-relevant model

12. Introduction Model Experiments Conclusion
Normalized Log-Likelihood Ratio
NLLR(D|R) = H(θD , θ¬R ) − H(θR , θD )
P(t|θR )
= P(t|θD ) log
P(t|θ¬R )
t∈V
(1 − δ1 )P(t|R) + δ1 P(t)
= P(t|θD ) log
(1 − δ2 )P(t|¬R) + δ2 P(t)
t∈V
• Measures how much better the relevant model can encode
events from the document model than the non-relevant
model
• If a term has a high probability of occurring in θR / θ¬R it is
rewarded / penalized

13. Introduction Model Experiments Conclusion
Normalized Log-Likelihood Ratio
NLLR(D|R) = H(θD , θ¬R ) − H(θR , θD )
P(t|θR )
= P(t|θD ) log
P(t|θ¬R )
t∈V
(1 − δ1 )P(t|R) + δ1 P(t)
= P(t|θD ) log
(1 − δ2 )P(t|¬R) + δ2 P(t)
t∈V
• Measures how much better the relevant model can encode
events from the document model than the non-relevant
model
• If a term has a high probability of occurring in θR / θ¬R it is
rewarded / penalized

14. Introduction Model Experiments Conclusion
Query Model
• Expanded query part
ˆ
P(t|θQ ) ∝ P(t|θD )P(θD |θR )
D∈R
where
NLLR(D|R)
P(θD |θR ) =
D NLLR(D |R)

15. Introduction Model Experiments Conclusion
Outline
Introduction
Model
Experiments
Conclusion

16. Introduction Model Experiments Conclusion
Experimental Setup
• Preprocessing
• Porter stemming
• Stopwords removed
• Training
• Optimize MAP on held-out set (odd-numbered topics)
• Sweep over free parameters
• λD , λQ
• δ1 for P(t|θR )
• δ2 for P(t|θ¬R )
• Submitted runs
• Used 10 terms with the highest P(t|θQ )
• met6: Non-relevant documents
• met9: Substitutes non-relevant model with collection

17. Introduction Model Experiments Conclusion
statMAP
A B C D E
met6 0.2289 0.2595 0.2750 0.2758 0.2822
met9 0.2289 0.2608 0.2787 0.2777 0.2810
indicates a statistically significant difference with the previous
set at the 0.01 level, tested using a Wilcoxon test

18. Introduction Model Experiments Conclusion
31 TREC Terabyte topics
MAP P5 P10
A 0.1364 0.2516 0.2452
met6 B 0.1726 0.3161 0.3194
met6 C 0.1682 0.3032 0.2968
met6 D 0.1746 0.3097 0.3065
met6 E 0.1910 0.3935 0.3645
met9 B 0.1769 0.3161 0.3194
met9 C 0.1699 0.3161 0.3032
met9 D 0.1738 0.4000 0.3710
met9 E 0.1959 0.2903 0.2871
/ indicates a statistically significant difference with the
baseline (set A) at the 0.05 / 0.01 level resp.

19. Introduction Model Experiments Conclusion
31 TREC Terabyte topics, set E
<num>814</num>
<title>Johnstown flood</title>
<desc>Provide information about the Johnstown Flood in Johnstown, Pennsylvania
</desc>
flood
johnstown
dam
club
AP P10
water
noaa baseline 0.3366 0.3000
gov met6 0.7853 1.0000
sir
www
time
0 0.125 0.250 0.375 0.500

20. Introduction Model Experiments Conclusion
31 TREC Terabyte topics, set E
<num>808</num>
<title>North Korean Counterfeiting</title>
<desc>What information is available on the involvement of the North Korean Government
in counterfeiting of US currency</desc>
north
korean
counterfeit
korea
AP P10
state
drug baseline 0.2497 0.6000
weapon met6 0.0096 0.0000
countri
nuclear
traffick
0 0.125 0.250 0.375 0.500

21. Introduction Model Experiments Conclusion
31 TREC Terabyte topics, set E
0.1925
0.1920
0.1915
0.1910
0.1905
MAP
0.1900
0.1895
0.1890
0.1885
0.1880
0.1875
0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
δ2

22. Introduction Model Experiments Conclusion
31 TREC Terabyte topics, set E
0.21
0.20
0.19
0.18
0.17
MAP
0.16
0.15
0.14
0.13
P(t|θQ ) = ˆ
(1 − λQ )P(t|Q) + λQ P(t|θQ )
0.12
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
λQ

23. Introduction Model Experiments Conclusion
31 TREC Terabyte topics, set E
0.2040
0.2020
0.2000
0.1980
MAP
0.1960
0.1940
0.1920
0.1900
5 15 25 35 45 55 65 75 85 95 105
Number of terms

24. Introduction Model Experiments Conclusion
Conclusion and Future Work
• Conclusion
• Modeled (non)relevant documents as separate models and
created a query model by sampling proportional to the
NLLR of these models
• Results improve over baseline
• Non-relevance information does not help significantly
• Future work
• Further analysis
• Compare with other, established RF methods
• Set/Estimate λQ based on relevance information
• amount
• confidence

25. Introduction Model Experiments Conclusion
Questions?
Edgar.Meij@uva.nl
http://www.science.uva.nl/~emeij